Async view trong Django dùng khi nào, và pitfall với ORM ra sao?

Từ Django 3.1, view có thể là async def.

Lợi rõ khi view chờ I/O nhiều — gọi external API, đọc nhiều endpoint song song, WebSocket.

import asyncio
import httpx
from django.http import JsonResponse

async def aggregate(request):
    async with httpx.AsyncClient() as client:
        a, b = await asyncio.gather(
            client.get('https://api.x/users'),
            client.get('https://api.y/posts'),
        )
    return JsonResponse({'users': a.json(), 'posts': b.json()})

Rào cản lớn nhất là ORM.

• Django ORM bản chất là sync — gọi Post.objects.filter(...) trong async view sẽ bị SynchronousOnlyOperation.
• Phải dùng async variant: await Post.objects.aget(...), await Post.objects.acreate(...), async for post in Post.objects.all(), hoặc wrap bằng sync_to_async(...).
• Driver Postgres mặc định (psycopg2) cũng là sync, nên Django chạy ORM trong threadpool chứ không phải event loop thực sự.
• Driver async thật sự là psycopg3 (Django 4.2+ support một phần).

async def view(request):
    posts = []
    async for post in Post.objects.all()[:10]:
        posts.append({'id': post.id, 'title': post.title})
    return JsonResponse({'posts': posts})

Đừng đổi mọi view sang async vì "nghe có vẻ nhanh hơn".

• Sync view dưới Gunicorn gthread xử lý 90% case rất ổn.
• Async chỉ thắng rõ khi đợi nhiều I/O song song; query DB tuần tự thì sync còn nhanh hơn vì không có overhead event-loop hop.

Since Django 3.1, views can be async def.

The win shows up when the view waits on I/O a lot — external APIs, fan-out reads, WebSocket.

import asyncio
import httpx
from django.http import JsonResponse

async def aggregate(request):
    async with httpx.AsyncClient() as client:
        a, b = await asyncio.gather(
            client.get('https://api.x/users'),
            client.get('https://api.y/posts'),
        )
    return JsonResponse({'users': a.json(), 'posts': b.json()})

The biggest pitfall is the ORM:
- The Django ORM is fundamentally sync. Calling Post.objects.filter(...) inside an async view → SynchronousOnlyOperation (Django catches it).
- You must use async variants: await Post.objects.aget(...), await Post.objects.acreate(...), async for post in Post.objects.all(). Or wrap with sync_to_async(...).
- The default Postgres driver (psycopg2) is sync — Django runs ORM in a threadpool, not a real event loop. The truly async driver is psycopg3 (partial support in Django 4.2+).

async def view(request):
    posts = []
    async for post in Post.objects.all()[:10]:
        posts.append({'id': post.id, 'title': post.title})
    return JsonResponse({'posts': posts})

Pitfall: Do not flip every view to async because "it sounds faster".

• A sync view under Gunicorn gthread handles 90% of cases very well.
• Async wins only when waiting on lots of parallel I/O; sequential DB queries are faster sync because there is no event-loop hop overhead.